Why in evaluation there are 1856 predictions

[fersebas]

Why there are 1856 predictions in evaluation (last 15 days) instead of 15246 - 288 + 1 = 1873? It depends on batch size?

Thank you for your answer.

Fernando

[DrownFish19]

The dataloader with drop_last =True will drop the last items which can not make another batch.

In the baseline, the batchsize is set at 32, and 1873/32 = 58.53125, 1856/32 = 58. Thus the last 17 items will be drop.

        if flag == 'test':
            shuffle_flag = False
            drop_last = True
        else:
            shuffle_flag = True
            drop_last = True

        ......
        data_loader = DataLoader(
            data_set,
            batch_size=self.args["batch_size"],
            shuffle=shuffle_flag,
            num_workers=self.args["num_workers"],
            drop_last=drop_last
        )

[zhoujingbo]

when the updated evaluation code will be released? It seems the batch-size matters in the current approach, which is not a good practice.

Hi, this is to update that we have updated the baseline

[phoenixes94]

def __read_data__(self):
    self.scaler = Scaler()
    df_raw = pd.read_csv(os.path.join(self.data_path, self.filename))
    border1s = [self.tid * self.total_size,
                self.tid * self.total_size + self.train_size - self.input_len,
                self.tid * self.total_size + self.train_size + self.val_size - self.input_len
                ]
    border2s = [self.tid * self.total_size + self.train_size,
                self.tid * self.total_size + self.train_size + self.val_size,
                self.tid * self.total_size + self.train_size + self.val_size + self.test_size
                ]
    border1 = border1s[self.set_type]
    border2 = border2s[self.set_type]

    df_data = df_raw
    if self.task == 'M':
        cols_data = df_raw.columns[self.start_col:]
        df_data = df_raw[cols_data]
    elif self.task == 'MS':
        cols_data = df_raw.columns[self.start_col:]
        df_data = df_raw[cols_data]
    elif self.task == 'S':
        df_data = df_raw[[self.tid, self.target]]

    # Turn off the SettingWithCopyWarning
    pd.set_option('mode.chained_assignment', None)
    df_data.replace(to_replace=np.nan, value=0, inplace=True)

    if self.scale:
        train_data = df_data[border1s[0]:border2s[0]]
        self.scaler.fit(train_data.values)
        data = self.scaler.transform(df_data.values)
    else:
        data = df_data.values
    self.data_x = data[border1:border2]
    self.data_y = data[border1:border2]
    self.raw_data = df_data[border1 + self.input_len:border2]

Self.tid * self.total_size + self.train_size + self.val_size - self.input_len Self.
data_x has the same dimension as self.data_y (16* 144,10), and self.raw_data (15* 144,10). Shouldn't the latitude of the predictions be int((24 * 6 * 16-288 + 1)/32)*32=2016?

[phoenixes94]

Looking forward to your reply!

[ramber1836]

As we mentioned before, we will not answer any questions related to the technical details in the Phase 3 competition.
Sorry for that.

[phoenixes94]

I just want to understand how the baseline code works, which is not a technical question about the structure of the model. Such problems in the past seem to be explained.